ODH and Princeton University Announce Machine Learning Pop Health Collaboration

Princeton University's Nassau Hall. Image is in the public domain.

This month, healthtech firm ODH, Inc. and Princeton University’s Department of Operations Research and Financial Engineering (ORFE) announced a partnership to develop machine learning techniques to better understand comorbidities and social determinants of health.

ODH will provide the healthcare expertise. Princeton will supply the “math power.”

Those are Adam Johnson’s words, not ours—he’s the Vice President of Product Development and Operations at ODH, and this week he and his company’s CEO, Michael Jarjour, spoke to Healthcare Analytics News™ about the arrangement. Both men think ORFE will be a natural complement to their goals.

READ>> First-of-its-Kind AI Tool for Diabetic Retinopathy Detection Approved by FDA

Since so much of healthcare data is actually cost data (think modern EHRs), and ORFE’s prowess is in finance and operations, they might be able to find ways to improve outcomes by bending the cost curve.

“How can you look at cost movement as an indicator of some potential problem? That’s a lot of what we’re trying to get out of the finance side,” Johnson added. Initially, according to Johnson, the nexus of disease and behavioral health will be the collaboration’s focus: How conditions like depression or bipolar disorder act as comorbidities to diabetes, or how independent variables like housing situation and substance use can contribute to things like liver failure.

To get those insights with machine learning, you need data, and lots of it. Johnson said that they were after a wide variety of information, ranging from the typical healthcare data that providers and payers curate (often out of their electronic health records) to regional census data. Jarjour added that some different sources will also play an important role.

“Many of the people in behavioral health programs are covered by Medicaid. And if you think about the poorer population, a large part of it is actually housed in jail,” he said. “Criminal justice data is extremely important for making decisions about treating a population so they don’t end up in that vicious cycle of being released and then getting picked right back up again by the jail system.”

Jarjour believes that those types of data will be key to ODH’s new models. “These are all triggers that could essentially take a patient from the normal cost structure to a significantly higher cost structure,” he said. “These independent variables we want our system to ingest and help us better identify what the challenges for the patient are, what’s the best way to provide appropriate care, and, by extension, how you can truly impact costs.”

Johnson said that the collaboration has goals in place through the end of the year, and they hope to produce a paper and convert their findings into a larger library to inform their future healthcare products. But the early work won’t necessarily be easy.

“The next generation of machine learning technologies for healthcare must confront unique technical challenges arising from this domain,” ORFE assistant professor Samory K. Kpotufe, said in a statement. He listed off some of the hurdles for predictive tools in healthcare, namely the tension between accuracy, interpretability, and patient privacy. “Confronting the challenges of machine learning for healthcare is bound to generate a rich set of research questions with potential impact beyond the healthcare domain.”

Another challenge of machine learning? Nailing down what people mean when they say the words. ODH’s release and our conversation with Johnson and Jarjour added to the growing library of differing definitions for the term. The ODH statement called it “an artificial intelligence technology”—some might argue that artificial intelligence is a machine learning technology—and the 2 executives each had a slightly different bend on the term.

Johnson even seemed at odds with that statement’s wording: “You get into the concept of artificial intelligence, and even semi-supervised machine learning could be that,” he said. “We consider machine learning to be when you have an output goal and an amount of information that’s hard for humans to process, so you can leverage the technology to get that output with some scale of prediction.”

To Jarjour, true machine learning occurs when the algorithms are unsupervised. He expressed disagreement with what healthcare venture capitalist Anya Scheiss told Healthcare Analytics News™ during an interview last month at HIMSS: “In her opinion, if you do an ‘if-and-then’ statement with Excel, that’s machine learning. I would say we have a very different opinion on that.”

But no matter how you slice it, Jarjour says ODH’s models are hungry for it.

“Our system allows for ingestion of any viable or interesting data set that can further improve our risk stratification,” he said. “Realistically, the goal there is to try to improve patient care.”

Related Coverage:

What's the Difference Between AI, Machine Learning, and Good, Old Analytics?

Demystifying Big Data and Machine Learning for Healthcare

A New Ethical Wrinkle for Medical Algorithms